The Costate-Free Revolution: A Simpler Path to Optimal Control

Opening

Every time an autonomous spacecraft fires its thrusters to reach a target orbit, every time a robotic arm adjusts its grip with minimum energy expenditure, every time an autonomous vehicle smoothly navigates a turn — somewhere behind the scenes, an optimal control algorithm is running. These algorithms answer a deceptively simple question: what is the best way to get from A to B while spending as little fuel, energy, or effort as possible?

For most of the past century, answering that question has required a mathematical workaround that nobody particularly loves. You want to minimize an objective — say, total fuel burn. But you also need your solution to obey the laws of physics. So the standard toolkit, codified in the mid-20th century through the work of Pontryagin and others, forces you to introduce a parallel set of hidden variables called costates. Think of them as Lagrange multipliers — bookkeeping ghosts that shadow every physical variable in your system and enforce the equations of motion as constraints on your optimization. They work. But they double the size of your problem, and they carry no direct physical intuition. You can't measure a costate. You can't easily initialize one. They exist purely to make the mathematics behave.

Now, Ossama Abdelkhalik and Aimar Negrete of Michigan Technological University are proposing something genuinely different (Abdelkhalik & Negrete, 2026). Their insight, elegant in retrospect, is this: you don't need costates at all — if you're willing to rethink what "the system" actually is.

The Science

The key move in this paper is conceptual before it is mathematical. Classical optimal control theory treats a dynamic system — a satellite, a robot arm, a chemical reactor — as a given, described by known equations of motion. Control forces enter as external inputs, and the job of the optimizer is to choose those inputs wisely. The physical laws and the optimization problem are handled in two separate layers, stitched together by costates.

Analytical mechanics, the branch of classical physics developed by Euler, Lagrange, and Hamilton in the 18th and 19th centuries, does something philosophically different. Rather than writing down equations of motion and then solving them, it derives those equations from a deeper principle: the minimization of a quantity called the action functional. The action is an integral over time of the difference between kinetic and potential energy — the Lagrangian, $\mathcal{L}=T-V$, where $T$ is kinetic energy and $V$ is potential energy. Extremize that integral, and out fall the equations of motion automatically. This is Hamilton's principle, and it is one of the most powerful unifying ideas in all of physics.

$$S={\int }_{t_0}^{t_f}\mathcal{L}(q,\dot{q},t)dt$$

What Abdelkhalik and Negrete do is ask: what if the control actuator isn't external to the system — what if it is part of the system? If you treat the actuator as a physical subsystem with its own energy contributions, you can add those energy terms directly to the action functional. Minimize the augmented action, and you get not just the equations of motion, but also equations governing the optimal control — all at once, from a single variational principle, with no costates required.

The authors present two distinct methods for constructing this modified action functional, offering different entry points depending on the structure of the problem at hand. Both methods rest on the same foundational idea: control effort has an energetic cost, and that cost can be written into the physics from the start, rather than bolted on afterward as an external constraint.

The paper grounds this abstract framework in a concrete case study — a canonical optimal control problem used widely as a benchmark — and demonstrates that the new approach produces physically valid, optimal solutions. The methodology draws on the Euler-Lagrange equations, the classical machinery for extremizing functionals, now applied to a richer action that includes control energy terms alongside the standard kinetic and potential contributions.

What They Found

The central result is structural: the proposed framework successfully derives the necessary conditions for optimal control — what mathematicians call the optimality conditions — without ever introducing a single costate variable (Abdelkhalik & Negrete, 2026). This is not a minor bookkeeping simplification. In standard optimal control, if your physical system has $n$ state variables, you need $n$ costate variables on top of them. For a spacecraft with six degrees of freedom — three positions, three velocities — that's twelve variables to track instead of six, with the costate initial conditions often unknown and requiring iterative guessing. The new approach keeps the variable count at $n$.

The two methods for constructing the modified action functional are shown to be mathematically consistent with each other and with the results that classical optimal control theory would produce for the same problem. This equivalence is critical: it means the new approach isn't an approximation or a heuristic — it is an exact reformulation. The solutions it yields are provably optimal in the same sense that Pontryagin's Maximum Principle guarantees optimality.

The case study demonstrates end-to-end: starting from a physical system description, augmenting the action functional with control energy terms, applying the Euler-Lagrange equations to the augmented functional, and arriving at a tractable system of differential equations whose solution gives both the state trajectory and the optimal control law simultaneously. The derivation is cleaner, the resulting equations are more symmetric, and the physical interpretation is more transparent — control and dynamics emerge from the same root principle.

Variable Count: Classical vs. Variational Optimal Control

Label	Value
Physical States (n)	6
Costates (standard OC)	6
Total Variables	12

Illustrative example for a 6-DOF system (3 positions, 3 velocities). Standard OC requires n additional costate variables; the new approach requires none.

Perhaps the most intellectually striking aspect of the finding is what it implies about the nature of costates themselves. From the variational mechanics perspective, costates appear to be artifacts of a particular mathematical formulation, not fundamental objects. When you embed the control problem into the physics correctly — by treating actuators as genuine physical subsystems — they dissolve. The paper suggests this is not just a computational convenience but a deeper insight into the relationship between control theory and mechanics.

Framework Comparison: Standard OC vs. Variational Mechanics Approach

Label	Value
Physical Interpretability	2
Problem Simplicity	2
Geometric Structure	2
Constraint Handling (current)	4
Established Tooling	5
Initialization Ease	2

Qualitative ratings (1–5) based on author descriptions and field context. Not empirical benchmarks.

Why This Changes Things

To appreciate why this matters, it helps to understand where optimal control is used today and where it is headed.

Modern autonomous systems — from spacecraft to surgical robots to self-driving vehicles — increasingly rely on real-time optimal control. The algorithms need to run fast, often on constrained hardware, often in environments where the problem structure changes dynamically. Every reduction in problem complexity translates directly into faster computation, smaller memory footprint, and more tractable numerical methods. The costate-elimination offered by this framework is therefore not just theoretically satisfying; it has practical teeth.

There is also a deeper issue with costates that the community has grappled with for decades: they are notoriously difficult to initialize. In the standard shooting method for solving two-point boundary value problems — the workhorse of spacecraft trajectory optimization — you must guess initial costate values, integrate forward, check whether boundary conditions are satisfied, and iterate. The sensitivity of the solution to those initial guesses is often severe, making convergence fragile. A framework that bypasses costates entirely sidesteps this initialization problem at its root.

The connection to analytical mechanics also opens a door that has been largely closed in control theory: the rich toolkit of symplectic geometry, geometric mechanics, and variational integrators developed over the past several decades. These tools have proven enormously powerful in physics and in the numerical simulation of Hamiltonian systems — they preserve energy and momentum in ways that standard integrators do not. Bringing optimal control into this geometric framework could enable a new class of structure-preserving optimal control algorithms, ones that are not just computationally efficient but physically faithful over long time horizons. For space mission design, where trajectories must be propagated over years, this matters enormously.

The approach also resonates with recent trends in physics-informed machine learning and differentiable simulation. When control and dynamics share a common variational origin, it becomes natural to differentiate through both simultaneously — exactly what modern automatic differentiation frameworks are built to do. This suggests pathways toward hybrid analytical-learned control architectures that could be more data-efficient and more interpretable than pure black-box approaches.

It is worth noting what this work does not claim. The paper presents a new theoretical framework and a proof-of-concept case study; it does not yet benchmark the approach against state-of-the-art solvers on a broad suite of problems, nor does it address all classes of constraints (inequality constraints on states and controls, for instance, are among the harder challenges in practical optimal control). These are known limitations that the authors acknowledge implicitly by framing this as a foundational contribution rather than a deployment-ready algorithm.

What's Next

The immediate open question is scope. The case study in the paper is deliberately chosen to be tractable and illustrative, but the real test of any new control framework is how it handles the ugly, high-dimensional, constrained problems that appear in engineering practice. Does the variational approach scale gracefully when $n$ is large? How does it handle state and control constraints — the kind that arise when a rocket can't exceed a maximum thrust or a joint can't bend past a physical limit? Extending the framework to handle inequality constraints will likely require additional machinery, possibly drawing on variational inequalities or contact mechanics.

There is also a question of numerical implementation. The Euler-Lagrange equations that emerge from the augmented action functional are a system of differential equations that must ultimately be solved numerically. The structure of that system — its stiffness, its boundary conditions, its sensitivity to perturbations — will determine how well existing numerical methods apply and what new solvers might be needed. The geometric structure of the variational formulation hints that variational integrators, which discretize the action functional directly rather than the equations of motion, could be the natural numerical companion to this framework.

Another rich direction is the treatment of stochastic systems. Classical optimal control has a well-developed stochastic branch — stochastic dynamic programming, the linear-quadratic-Gaussian regulator — and it would be natural to ask whether a variational mechanics approach can be extended to systems with noise. Stochastic variational mechanics exists as a mathematical field, and connecting it to this control framework could open new territory.

Perhaps most speculatively: if control effort can be encoded as energy in the action functional, what does that suggest about the relationship between control theory and thermodynamics? The minimum-effort control problem has echoes of least-dissipation principles in non-equilibrium thermodynamics. Whether this parallel is merely aesthetic or points to something deeper — a unified framework for energy-optimal processes across physics and engineering — is an open question worth pursuing.

What Abdelkhalik and Negrete have done, at minimum, is demonstrate that a costate-free formulation of optimal control is not just possible but mathematically coherent and physically motivated. That proof of concept, in a field where foundational frameworks have been stable for decades, is itself a significant achievement. The path from here to widespread adoption will require careful work on numerical methods, constraint handling, and validation on complex systems. But the conceptual door is open, and the view through it is genuinely new.

For anyone who has spent time wrestling with the sensitivity of costate initialization, the opacity of adjoint variables, or the gap between elegant theory and practical solver behavior, this paper reads as an invitation: there may be a simpler underlying structure to optimal control, one that the field has been circling around for a century without quite finding. Analytical mechanics found the equations of motion by asking what nature minimizes. Now, the same question is being asked of control itself.